Multi-task model training method, electronic equipment, storage medium and program product

Abstract

The embodiment of the invention provides a multi-task model training method, electronic equipment, a storage medium and a program product. The method comprises the steps of determining a standardized performance data set and a standardized target data set corresponding to each group of time periods on the basis of historical time periods, corresponding to each piece of standardized performance data, of each performance index and future time periods corresponding to each piece of standardized target data; based on a preset multi-dimensional time feature, determining an input time feature vector of the standardized performance data set corresponding to each group of time periods and a target time feature vector of the standardized target data set, and further determining a plurality of time period data sets corresponding to the plurality of group of time periods; and a prediction model is reconstructed, correction processing is performed on each time period data set to obtain a plurality of target time period data sets corresponding to the plurality of group time periods, and training processing is performed on the initial multi-task model through the plurality of target time period data sets to obtain a target multi-task model. The method is used for achieving the effect of improving the accuracy of multi-task model training.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a multi-task model training method, electronic device, storage medium, and program product. Background Technology

[0002] With the popularization of 5G networks and the increasing complexity of network structures, data characteristics in mobile communication networks exhibit high dimensionality, multiple dimensions, strong periodicity, and significant regional heterogeneity.

[0003] In existing technologies, when training a multi-task model using sample data, missing sample data can be imputed, and all the supplemented data features can be directly concatenated. A shared encoder is then used to complete the multi-task processing to obtain multi-task training samples, which are then used to train the model. However, directly concatenating data features from different time periods can cause crosstalk between these periods, leading to compression of the predicted amplitude of weakly dimensional indicators and resulting in lower accuracy in training the multi-task model. Summary of the Invention

[0004] This application provides a multi-task model training method, electronic device, storage medium, and program product to improve the accuracy of multi-task model training.

[0005] In a first aspect, embodiments of this application provide a multi-task model training method, including:

[0006] Obtain multiple standardized input values ​​and multiple standardized predicted values ​​corresponding to the initial multi-task model. The standardized input values ​​include multiple standardized performance data corresponding to each performance indicator among multiple performance indicators, and the standardized predicted values ​​include multiple standardized target data corresponding to each performance indicator.

[0007] Based on the historical time periods corresponding to each standardized performance data and the future time periods corresponding to each standardized target data, the standardized performance dataset and standardized target dataset corresponding to each group of time periods are determined. The group of time periods is the time period in which the historical time period and the future time period have a temporal correlation relationship.

[0008] Based on preset multidimensional time features, the input time feature vector of the standardized performance dataset corresponding to each group of time periods and the target time feature vector of the standardized target dataset are determined.

[0009] Based on the standardized performance dataset, input time feature vector, standardized target dataset and target time feature vector corresponding to each group of time periods, determine multiple time period datasets corresponding to the multiple groups of time periods;

[0010] By reconstructing the prediction model and correcting the datasets for each time period, multiple target time period datasets corresponding to the multiple time periods are obtained. The initial multi-task model is then trained using the multiple target time period datasets to obtain the target multi-task model.

[0011] Secondly, embodiments of this application provide a multi-task model training apparatus, including an acquisition module, a first determination module, a second determination module, a third determination module, a correction processing module, and a training module:

[0012] The acquisition module is used to acquire multiple standardized input values ​​and multiple standardized prediction values ​​corresponding to the initial multi-task model. The standardized input values ​​include multiple standardized performance data corresponding to each performance indicator among multiple performance indicators, and the standardized prediction values ​​include multiple standardized target data corresponding to each performance indicator.

[0013] The first determining module is used to determine the standardized performance dataset and standardized target dataset corresponding to each group of time periods based on the historical time periods corresponding to each standardized performance data and the future time periods corresponding to each standardized target data, wherein the group of time periods is a time period in which the historical time period and the future time period have a temporal correlation relationship;

[0014] The second determining module is used to determine, based on preset multi-dimensional time features, the input time feature vector of the standardized performance dataset corresponding to each group of time periods, and the target time feature vector of the standardized target dataset;

[0015] The third determining module is used to determine multiple time period datasets corresponding to the multiple groups of time periods based on the standardized performance dataset, input time feature vector, standardized target dataset and target time feature vector corresponding to each group of time periods;

[0016] The correction processing module is used to correct the datasets of each time period by reconstructing the prediction model, so as to obtain multiple target time period datasets corresponding to the multiple groups of time periods.

[0017] The training module is used to train the initial multi-task model using the multiple target time period datasets to obtain the target multi-task model.

[0018] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0019] The memory stores computer-executed instructions;

[0020] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0021] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0022] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0023] The multi-task model training method, electronic device, storage medium, and program product provided in this application determine the time feature vector corresponding to each time period based on standardized input values ​​and standardized predicted values. At the same time, the channel-independent modeling method is used to eliminate feature crosstalk between heterogeneous indicators, which can reduce feature crosstalk between performance indicators in different time periods, improve the model's adaptability to data with high missing rate, and thus improve the accuracy of multi-task model training. Attached Figure Description

[0024] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0025] Figure 1 A schematic diagram illustrating the application scenarios provided in this application;

[0026] Figure 2 A flowchart illustrating a multi-task prediction method provided in this application;

[0027] Figure 3 A flowchart illustrating another multi-task model training method provided in this application;

[0028] Figure 4 A schematic diagram of the architecture of a multi-task model training method provided in this application;

[0029] Figure 5 A schematic diagram of the architecture of a multi-task model provided in this application;

[0030] Figure 6 A schematic diagram of the structure of a multi-task model training device provided in this application;

[0031] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0032] The accompanying drawings have illustrated specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to specific embodiments. Detailed Implementation

[0033] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0034] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with relevant laws, regulations and standards, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0035] Figure 1 A schematic diagram illustrating the application scenarios provided in this application. Please refer to [link / reference]. Figure 1 In the operation and maintenance scenario of a 5G cellular network 100, multiple base stations 101 and network monitoring equipment 102 can be included. The network monitoring equipment 102 can monitor the performance data corresponding to multiple performance indicators of each base station 101 in real time. The performance indicators can be signal strength, traffic load, number of user connections, etc.

[0036] Network monitoring equipment 102 needs to process high-missing-rate data (such as complete missing data or sudden breakpoints caused by equipment failure) under the complex topology of multiple base stations 101, and train an initial multi-task model using historical performance data to obtain a target multi-task model, and then use the target multi-task model to predict multi-task prediction results. For example, during holidays or sudden events, network traffic may surge, and network monitoring equipment 102 needs to quickly adjust resource allocation based on multi-task prediction results.

[0037] In existing technologies, when training a multi-task model using sample data, missing sample data can be imputed, and all the supplemented data features can be directly concatenated. A shared encoder is then used to complete the multi-task processing to obtain multi-task training samples, which are then used to train the model. However, directly concatenating data features from different time periods can cause crosstalk between performance metrics, leading to compression of the predicted amplitude of weakly dimensional metrics and resulting in lower accuracy in training the multi-task model.

[0038] The multi-task model training method provided in this application obtains multiple standardized input values ​​and multiple standardized predicted values ​​corresponding to an initial multi-task model. The standardized input values ​​include multiple standardized performance data corresponding to each performance indicator, and the standardized predicted values ​​include multiple standardized predicted values ​​corresponding to each performance indicator. Based on the historical time periods corresponding to each standardized performance data and the future time periods corresponding to each standardized predicted value, the standardized performance dataset corresponding to the historical time periods and the standardized target dataset corresponding to the future time periods are determined. Based on preset multi-dimensional time features, the input time feature vector of the standardized performance dataset corresponding to the historical time periods and the target time feature vector of the standardized target dataset corresponding to the future time periods are determined, thus determining multiple time period datasets corresponding to multiple time periods. By reconstructing the prediction model and correcting the time period datasets, multiple target time period datasets are obtained. The initial multi-task model is then trained using these multiple target time period datasets to obtain the target multi-task model.

[0039] In the above execution process, based on the standardized input values ​​and standardized predicted values, the time feature vectors corresponding to each time period are determined. At the same time, the channel-independent modeling method is used to eliminate feature crosstalk between heterogeneous indicators, which can reduce feature crosstalk between performance indicators in different time periods, improve the model's adaptability to data with high missing rates, and thus improve the accuracy of multi-task model training.

[0040] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0041] Figure 2 This is a flowchart illustrating a multi-task prediction method provided in this application. Please refer to [link / reference]. Figure 2 The method includes:

[0042] S201. Obtain multiple standardized input values ​​and multiple standardized prediction values ​​corresponding to the initial multi-task model.

[0043] The execution subject of this application embodiment can be a network monitoring device or a multi-task prediction device installed in the network monitoring device. The multi-task prediction device can be implemented by software or by a combination of software and hardware.

[0044] Multiple standardized input values ​​correspond one-to-one with multiple standardized predicted values, and a standardized input value and its corresponding standardized predicted value constitute a standardized sample.

[0045] The standardized input values ​​include multiple standardized performance data corresponding to each performance indicator among multiple performance metrics, and the standardized predicted values ​​include multiple standardized target data corresponding to each performance indicator.

[0046] For example, the standardized input value is multiple performance data of each performance indicator of the base station on the previous day, and the standardized prediction value is multiple performance data of each performance indicator of the base station on the next day.

[0047] Among them, multiple standardized input values ​​can be used Indicators suggest that multiple standardized forecasts can be used Instructions, among which, The number of samples for multiple standardized samples. The input sample length (i.e., the number of time steps in the historical period) for multiple standardized performance data. The number of performance indicators, The length of the output value of multiple standardized target data in the standardized forecast (i.e., the number of time steps in the forecast period).

[0048] There are many types of performance metrics, which can cover signal strength, traffic load, number of user connections, etc., and the units of measurement of different metrics vary greatly (such as percentage, absolute value, number of events per unit time, etc.).

[0049] For example, if the performance metric is signal strength, then the multiple standardized performance data points are the historical signal strength data corresponding to 96 time periods of the previous day, i.e. Multiple standardized target data can be used to generate future signal strength data for 96 time periods on the following day, i.e. .

[0050] S202. Based on the historical time periods of each standardized performance data corresponding to each performance indicator and the future time periods of each standardized target data, determine the standardized performance dataset and standardized target dataset corresponding to each set of time periods.

[0051] Each historical period corresponds to a set of standardized performance data for each performance indicator, and each future period corresponds to a set of standardized target data for each performance indicator.

[0052] There is a temporal relationship between historical time periods and future time periods. Based on the temporal relationship, the standardized performance dataset and standardized target dataset corresponding to each historical-future time period pair can be extracted.

[0053] The standardized performance dataset includes standardized performance data corresponding to each performance indicator in the historical period, and the standardized target dataset includes standardized target data corresponding to each performance indicator in the future period.

[0054] For example, there are 3 performance indicators (performance indicators 1-3), the historical period consists of 4 consecutive time steps (periods 1-4, corresponding to L=4), and the future period consists of 4 consecutive time steps (periods 5-8, corresponding to H=4). The historical period and the future period are associated in the following time sequence: "period 1 → period 5, period 2 → period 6, period 3 → period 7, period 4 → period 8", as shown in Tables 1 and 2.

[0055] Therefore, the standardized performance dataset 1 corresponding to historical period 1 is {standardized performance data 11, standardized performance data 21, standardized performance data 31}, and the corresponding standardized target dataset 1 is {standardized target data 15, standardized target data 25, standardized target data 35}.

[0056] The standardized performance dataset 2 corresponding to historical period 2 is {standardized performance data 12, standardized performance data 22, standardized performance data 32}, and the corresponding standardized target dataset 2 is {standardized target data 16, standardized target data 26, standardized target data 36}.

[0057] The standardized performance dataset 3 corresponding to historical period 3 is {standardized performance data 13, standardized performance data 23, standardized performance data 33}, and the corresponding standardized target dataset 3 is {standardized target data 17, standardized target data 27, standardized target data 37}.

[0058] The standardized performance dataset 4 corresponding to historical period 4 is {Standardized Performance Data 14, Standardized Performance Data 24, Standardized Performance Data 34}. The corresponding standardized target dataset 4 is {Standardized Target Data 18, Standardized Target Data 28, Standardized Target Data 38}.

[0059] Table 1

[0060]

[0061] Table 2

[0062]

[0063] S203. Based on the preset multi-dimensional time features, determine the input time feature vector of the standardized performance dataset corresponding to each time period, and the target time feature vector of the standardized target dataset.

[0064] In some embodiments, multiple time feature rules corresponding to preset multidimensional time features can be obtained; through multiple time feature rules, the time series information of the standardized performance dataset corresponding to each historical period is processed by feature extraction to obtain the input time feature vector corresponding to each historical period; through multiple time feature rules, the time series information of the standardized target dataset corresponding to each future period is processed by feature extraction to obtain the target time feature vector corresponding to each future period.

[0065] Multiple time feature rules can include multiple periodic item rules and multiple business time period rules. The periodic item rules are used to extract hourly periodic features (sine / cosine coding) and 15-minute granularity periodic features (sine / cosine coding). The business time period rules are used to extract business scenario identification features (such as one-hot coding of working hours, morning and evening peak hours, and nighttime off-peak hours).

[0066] The input time feature vector is used to indicate multiple time features of the standardized performance dataset for its corresponding historical period; the target time feature vector is used to indicate multiple time features of the standardized target dataset for its corresponding future period.

[0067] Multiple standardized input values The corresponding input multidimensional time features are used Indicators, multiple standardized forecasts The corresponding output multidimensional time features are used Instructions. Among them, The number of samples for multiple standardized samples. The input sample length for multiple standardized performance data. To predefine the dimensions of multidimensional time features, The length of the output value for multiple standardized target data.

[0068] In this application, by pre-setting multi-dimensional time features, the time features of each standardized performance dataset and each standardized target dataset can be extracted, providing a stable time series benchmark for the model, supporting generalization ability across time zones and time periods, and thus improving the accuracy of model training.

[0069] S204. Based on the standardized performance dataset, input time feature vector, standardized target dataset and target time feature vector corresponding to each group of time periods, determine multiple time period datasets corresponding to multiple groups of time periods.

[0070] In some embodiments, an observation mask matrix can be generated; the observation mask matrix is ​​used to mask the standardized performance datasets corresponding to each time period to obtain the masked performance datasets corresponding to each time period; based on the masked performance datasets corresponding to each time period, the input time feature vector, and the standardized target dataset and target time feature vector, the time period datasets corresponding to each time period are determined.

[0071] The observation mask matrix can be used to indicate the locations in the standardized performance dataset corresponding to each time period that need to be masked. By simulating missing data that may exist in real-world scenarios (such as missing columns or sudden breakpoints caused by equipment failure), the model learns the ability to reconstruct missing data during the training phase, thereby improving the reliability and generalization of the model training in real-world scenarios.

[0072] The observation mask matrix can include real missing masks, full-column composite masks, and time-slice composite masks. Each mask is generated independently and then combined by element-wise multiplication (⊙) to obtain the final observation mask matrix.

[0073] The true missing data mask is used to mark the location of naturally missing data caused by equipment failure, transmission interruption, etc. A mask value of 0 indicates that the data is missing, and 1 indicates that the data is valid.

[0074] The entire column synthesis mask is used to simulate a scenario where a certain performance indicator is completely missing. A portion of the performance indicator columns are randomly selected according to a preset probability, and the mask value of all time steps of that column is set to 0.

[0075] The time-slice synthesis mask is used to simulate a scenario where a certain performance indicator is missing consecutive time steps. The performance indicator column, consecutive time slice length and start position are randomly selected according to a preset probability, and the mask value of the corresponding time slice is set to 0.

[0076] In this context, the mask bit corresponding to the time feature vector is always set to 1, meaning that the time features are always visible to the model, providing a stable benchmark for time series modeling.

[0077] For example, if the "traffic load" data for a certain historical period is naturally missing due to equipment failure, the corresponding position of the actual missing mask is 0, and the model needs to learn and reconstruct at that position; to simulate the extreme scenario of a performance indicator being completely missing, the entire column of the composite mask can be used to set all time steps of the indicator column to 0, forcing the model to rely on other indicators and time features to complete the reconstruction; to simulate continuous missing data caused by sudden breakpoints, the time slice composite mask can be used to set three consecutive time steps of the "traffic load" indicator to 0, training the model's time-series completion capability.

[0078] By allowing the model to directly learn the reconstruction task of missing locations during training, while simultaneously optimizing the prediction task based on an unmasked, standardized target dataset, the model can optimize both reconstruction and prediction tasks in end-to-end training, achieving synergistic optimization of the two tasks.

[0079] In this application, by generating observation mask matrices with multiple types of combinations, missing data modeling and prediction tasks are organically combined, which solves the problem of the separation between missing data modeling and prediction objectives in traditional methods and significantly improves the prediction accuracy of the model in real missing data scenarios.

[0080] S205. By reconstructing the prediction model, the datasets for each time period are corrected to obtain multiple target time period datasets corresponding to multiple time periods.

[0081] The reconstruction prediction model is an end-to-end joint training architecture, including a channel-independent encoder, a reconstruction head, an autoregressive prediction head, a column gain correction unit, and a joint loss calculation unit.

[0082] The channel-independent encoder can receive a masked performance dataset and an input time feature vector from a time-segment dataset. It first concatenates each performance metric with the same time feature vector, then rearranges them into a "..." structure. The hidden sequence is output by extracting the temporal features of various performance indicators in a channel-independent form, through a shared multi-head self-attention and residual feedforward network. ,in, For encoder dimensions.

[0083] The reconstructed head can be based on hidden sequences We employ time-progressive MLP decoding to reconstruct the mask locations (missing data) in the standardized performance dataset, outputting a standardized reconstruction result with the same dimensions as the input. It is used to fill in missing data for historical periods.

[0084] Autoregressive prediction heads can be based on the same hidden sequence. By sharing an autoregressive decoding structure, it learns the mapping from historical time-series features to future time-series features and outputs preliminary standardized prediction results. .

[0085] The column gain correction unit can introduce learnable gain parameters per performance index. The initial standardized prediction results are then subjected to amplitude alignment correction to obtain the final standardized prediction results. ( (Element-wise multiplication) is used to address the problem of compressed amplitude predictions for different dimensional indices.

[0086] The joint loss calculation unit is used to calculate the four losses that constitute the joint objective (reconstruction loss). Predicting losses Column gain regularization loss Temporal consistency loss This provides a supervision signal for updating model parameters.

[0087] The specific processing flow is as follows: Input the time period dataset (mask performance dataset) Input time feature vector Standardized target dataset Target time feature vector Observation mask The temporal features are extracted by the independent encoder of the channel, and the intermediate features between the reconstruction branch and the prediction branch are output synchronously; the reconstruction head pair The mask positions are padded to generate The prediction head is generated after column gain correction. ;Will and splicing and By splicing, the complete reconstructed tensor is obtained. and prediction tensor By combining the calculation results of the joint loss function, the "reconstruction tensor R, prediction tensor F, and joint loss value L" are packaged to obtain the target time period dataset corresponding to each time period.

[0088] S206. Using multiple target time period datasets, train the initial multi-task model to obtain the target multi-task model.

[0089] During training, multiple target time-period datasets are input, including the reconstructed tensor R, the predicted tensor F, the joint loss value L, and the normalized target dataset. Observation mask Column gain parameters .

[0090] The optimization objective is to minimize the joint loss function. By co-supervising the reconstruction and prediction tasks, the model's ability to complete missing data and predict future time series can be optimized simultaneously.

[0091] The parameters of the channel-independent encoder, reconstruction head, autoregressive prediction head, and column gain correction unit of the initial multi-task model are initialized; the column gain parameter g is initialized to a value in the range of 8 to 16 to ensure that the predicted amplitude of the weak dimension index is not over-compressed.

[0092] During batch iterative training, the dataset for the target time period can be divided into training batches according to the batch size B (default 16), and all training data can be iterated through. For each batch, the joint loss value L is calculated forward, and the gradients of each model parameter are calculated using the backpropagation algorithm (chain rule). The AdamW optimizer is used to update the model parameters, and the learning rate is set. The weight decay coefficient is (Suppress overfitting); Perform gradient pruning (threshold=1.0) to avoid gradient explosion causing model training instability; Optional EMA (exponential moving average, coefficient=0.995) can be enabled to improve model generalization performance and prediction stability.

[0093] The training epochs are ≤120 to avoid overfitting due to overtraining; a validation set (10%~20% of the model's performance) is introduced to monitor model performance, with the "validation set prediction loss" as the benchmark. "This is the early stopping criterion. If the loss does not decrease after 10 consecutive rounds of verification, training will stop and the historical best parameters will be retained. The loss weights are fixed with the default configuration: reconstruction loss weight α=0.05, prediction loss weight γ=0.20, column gain regularization weight λ=0.0 (default off), and temporal consistency loss weight β=0.0 (default off). Adjustments can be made according to the actual scenario."

[0094] Additionally, regularization terms can be added to suppress excessive offsets:

[0095]

[0096]

[0097] in, The square of the L2 norm constrains the degree to which the gain parameter deviates from 1, thus preventing excessive amplification of the predicted amplitude. To sum the absolute values ​​of each element, To prevent the removal of zero minimum values, the consistency of the historical last position and the predicted first position is constrained.

[0098] After training, save the complete parameters of the historical best model (including the weights and biases of the encoder, reconstruction head, prediction head, and column gain correction unit); simultaneously save the key configurations fixed during training: normalizer parameters (scaler.pkl), feature column order (features[]), time feature dimension T=7, scaling factor s, schema version ( ), forming a self-contained model checkpoint;

[0099] The final target multi-task model must meet the following requirements: it should support input standardized performance data and time features, output the reconstruction results and prediction results of the standardized space, and be able to be restored to the original physical dimensions after inverse transformation. It should also have the ability to handle complex scenarios such as missing columns and missing consecutive time slices.

[0100] After the model training is completed, the core performance is verified through the test set: the MAE (mean absolute error) and RMSE (root mean square error) of the prediction task are more than 30% lower than the baseline model; in the missing data reconstruction task with a missing rate of 30%~50%, the relative error between the reconstructed data and the real data is ≤10%; the single-cell inference latency P50 is <20ms (based on A100 graphics card, K≈50 performance indicators), which meets the requirements of real-time operation and maintenance scenarios.

[0101] The multi-task model training method provided in this application can determine the standardized performance dataset corresponding to the historical period and the standardized target dataset corresponding to the future period based on the historical time period of each standardized performance data corresponding to each performance index and the future time period corresponding to each standardized predicted value. Based on preset multi-dimensional time features, it determines the input time feature vector of the standardized performance dataset corresponding to the historical period and the target time feature vector of the standardized target dataset corresponding to the future period, thus determining multiple time period datasets corresponding to multiple time periods. By reconstructing the prediction model and correcting the time period datasets, multiple target time period datasets are obtained. The initial multi-task model is then trained using these multiple target time period datasets to obtain the target multi-task model. Based on the standardized input values ​​and standardized predicted values, the time feature vectors corresponding to each time period are determined. Simultaneously, a channel-independent modeling approach is used to eliminate feature crosstalk between heterogeneous indices, which can reduce feature crosstalk between performance indices in different time periods, improve the model's adaptability to data with high missing rates, and thus improve the accuracy of multi-task model training.

[0102] Figure 3 A flowchart illustrating another multi-task model training method provided in this application. Please refer to... Figure 3 In this embodiment Figure 2 Based on the examples, the multi-task model training method is described in detail, which includes:

[0103] S301. Obtain multiple initial input values ​​corresponding to multiple standardized input values ​​and multiple initial prediction values ​​corresponding to multiple standardized prediction values.

[0104] It can obtain multiple initial input values ​​and multiple initial prediction values ​​from different sources and of different magnitudes.

[0105] The initial input values ​​correspond to the initial performance data corresponding to each performance characteristic of the base station during historical time periods (i.e., The initial predicted value corresponds to the initial target data of the base station for future time periods (i.e., The two must correspond one-to-one in historical-future time sequence and carry complete timestamp information and cell identifier (CGI).

[0106] The initial input values ​​include multiple initial performance data corresponding to each performance indicator, and the initial prediction values ​​include multiple initial target data corresponding to each performance indicator. The performance indicators cover signal strength, traffic load, number of user connections, etc., and data quality control must be completed in advance (deleting columns with high missing rates, filtering poor samples, and aligning timestamps with 15-minute granularity).

[0107] S302. Determine multiple intermediate input values ​​corresponding to multiple initial input values ​​and multiple intermediate prediction values ​​corresponding to multiple initial prediction values.

[0108] Initial performance data from different sources and of varying magnitudes are converted into a unified standardized format. Robust statistical estimation based on the entire network data is used, and standardization is performed using the median and interquartile range to eliminate data distribution differences between different regions and equipment manufacturers.

[0109] Please refer to the following standardized formula:

[0110]

[0111]

[0112] in, The standardized input value for the t-th time-series dimension corresponding to the c-th performance feature of the b-th training sample; The initial performance data for the c-th performance feature of the b-th training sample in the t-th time-series dimension; The standardized target data for the c-th performance feature of the b-th training sample is the standardized target data for the t-th time-series dimension. This is the initial target data for the t-th time-series dimension corresponding to the c-th performance feature of the b-th training sample.

[0113] It is worth noting that, and It can be a global robust statistic determined by training samples during the training phase, and it needs to be estimated after splicing and aggregating the sample dimension and the time dimension to ensure the stability and representativeness of the statistic.

[0114] You can refer to the following formula to determine it. :

[0115]

[0116] in, Indicates all initial performance data for the c-th performance characteristic. The initial target data indicating the c-th performance characteristic are calculated by merging the two types of data to ensure that the standardized benchmark covers the complete distribution of the input and the target.

[0117] You can refer to the following formula to determine it. :

[0118]

[0119] in, Combined data used to indicate the c-th performance feature among all training samples The corresponding upper quartile (i.e., the 75th percentile); Combined data used to indicate the c-th performance feature among all training samples The corresponding lower quartile (i.e., the 25th percentile).

[0120] The intermediate input value is a normalized quantity without a scaling factor. (Right now The intermediate predictions are standardized values ​​without scaling factors. (Right now Both are transitional forms from raw data to final standardized data, used for subsequent uniform scaling processing.

[0121] S303. Scale the multiple intermediate input values ​​and multiple intermediate predicted values ​​respectively to obtain multiple standardized input values ​​and multiple standardized predicted values.

[0122] Based on uniform scaling factor ( ,default ), for intermediate input values and intermediate forecast values Element-wise scaling is performed to obtain the final standardized input value. and standardized target data :

[0123]

[0124]

[0125] The final output is a normalized tensor. (Historical Input) (Future goals), where B is the number of samples, L=96 (historical time steps, 15 minutes / step), H=96 (future time steps, 15 minutes / step), and K is the number of performance indicators.

[0126] For any decoded normalized spatial output (Right now (The reconstructed / predicted results output by the model) should be restored to their original physical dimensions using the following steps:

[0127]

[0128] The inverse transformation sequence is: divide by s → denormalize (restore the unscaled normalized quantity) → physical pruning (limit the result to a reasonable range for business, such as non-negative signal strength and traffic load not exceeding the device limit).

[0129] Let Z be the unscaled normalized quantity, and Ẑ be the quantity with a uniform scaling factor s: for any primitive scalar (can be) or (at a certain location)

[0130]

[0131]

[0132]

[0133] Key parameters in the standardization process Persist to the scaler.pkl file, where, This is a global median array (length K) for each performance metric. The array represents the global interquartile range (length K) for each performance metric; cols is the list of column names and their order for the performance metrics (aligned with the model input); Version is the schema version of the normalizer (to avoid version compatibility issues).

[0134] S304. Based on the historical time periods of each standardized performance data corresponding to each performance indicator and the future time periods of each standardized target data, determine the standardized performance dataset and standardized target dataset corresponding to each set of time periods.

[0135] The execution process of S304 can be found in the execution process of S202, and will not be repeated here.

[0136] S305. Obtain multiple time feature rules corresponding to preset multidimensional time features.

[0137] Multiple time-feature rules can include multiple periodic item rules and multiple business time period rules.

[0138] Among them, multiple periodic item rules can be 4-dimensional periodic features, and multiple business time period rules can be 3-dimensional business time period identifiers. That is, the preset multi-dimensional time features are 7-dimensional. All rules need to be configured based on the actual operation and maintenance needs of the operation side to ensure that the time features are strongly correlated with the business scenarios.

[0139] Periodicity rules are used to capture the natural periodicity of time series data (daily cycle, 15-minute granularity cycle), avoid the model's dependence on absolute timestamps, and improve the model's ability to generalize across dates; business time period rules are used to inject domain business knowledge, strengthen the model's ability to model the differences between working hours / peak hours / off-peak hours, and improve prediction and reconstruction accuracy.

[0140] S306. Through multiple time feature rules, the time series information of the standardized performance dataset corresponding to each historical period is processed by feature extraction to obtain the input time feature vector corresponding to each historical period.

[0141] In some embodiments, feature processing can be performed on time periods based on multiple periodic item rules to obtain time feature values ​​corresponding to each periodic item rule; feature processing can be performed on time periods based on multiple business time period rules to obtain business time period feature values ​​corresponding to each business time period rule; the time feature values ​​corresponding to each periodic item rule and the business time period feature values ​​corresponding to each business time period rule are concatenated to obtain the input time feature vector corresponding to the standardized performance data.

[0142] The true start timestamps corresponding to each standardized performance data can be processed using time coding units: Let the true start hour of the sequence be... The value ranges from 0 to 23, extracted directly from the data timestamp; the time granularity is 15 minutes. Indexing performance data in time series .

[0143] The following formula is for calculating time granularity mapping:

[0144]

[0145]

[0146] Index 15 minutes Convert to hourly offset Ensure the hourly cycle remains within the range of 0 to 23. Convert the starting hour to a 15-minute index. Ensure that the cycle is completed within 96 15-minute time steps per day.

[0147] The following is the formula for calculating the 4-dimensional periodic term feature:

[0148]

[0149]

[0150]

[0151]

[0152] and It uses hourly cycle characteristics to capture 24-hour daily cycle patterns. and It features 15-minute cycle characteristics, capturing fine-grained cycle patterns across 96 time steps within a day.

[0153] For business time period identifiers (3D one-hot), set the business time period. Configured by operational rules (e.g., S1 = working hours, S2 = morning and evening peak hours, S3 = nighttime off-peak hours), then This corresponds to one-hot encoding.

[0154] Finally, the vector of multidimensional temporal features for a single time step (i.e., a time period) is:

[0155]

[0156] in, If and only if Otherwise, it is 0; Otherwise, it is 0; If and only if Otherwise, it is 0.

[0157] After performing the above feature extraction on all training samples, the output is the input time feature tensor corresponding to each historical time period.

[0158] in, Let B be the time feature dimension, and let B be the number of samples. This represents the length of the historical sequence.

[0159] Before feature extraction, it is necessary to verify the integrity of the timestamps in the standardized performance dataset: ensure that the historical sequence timestamps of each sample are continuous without any breaks, and that the starting timestamp format is uniform (such as YYYY-MM-DD HH:MM:SS) to avoid deviations in the calculation of periodic features due to abnormal timestamps.

[0160] S307. By using multiple time feature rules, the time series information of the standardized target dataset corresponding to each future time period is processed to extract features, and the target time feature vector corresponding to each future time period is obtained.

[0161] The true start hour of the future sequence It is derived from the end time of the corresponding historical sequence, that is Time index of future sequence ( ).

[0162] Using the same periodic term calculation, business time period encoding, and vector concatenation logic as S306, generate the target time feature tensor corresponding to each future time period. Where H=96 is the future prediction step size, and B is the sample size. For time features.

[0163] Save the time feature rules (including the calculation formula for periodic items, the configuration of business time periods, and the time granularity parameters) together with the normalizer and model weights. The rules must be strictly reused during the inference phase to ensure that the time feature calibers of training and inference are completely aligned.

[0164] S308. Generate the observation mask matrix.

[0165] The observation mask matrix is ​​used to indicate the mask position in the normalized performance dataset corresponding to each time period.

[0166] The following are the semantics and independence definitions of the mask values: This indicates an observation point that is "available for network use" (data valid / visible); The mask bit indicating that the data is "not visible to the network" (data is missing / reconstruction is required) is only used for the supervision calculation of subsequent reconstruction loss;

[0167] Masks are sampled independently for each sample (the mask distribution of different samples does not affect each other); the synthesis strategies for "complete column missing" and "time slice missing" are independent of each other; the mask bits corresponding to the time features are always set to 1 (ensuring that the time features are always visible to the model, providing a stable benchmark for time series modeling).

[0168] When constructing the observation mask matrix, the input data is a normalized performance dataset. (Where the actual missing locations are marked as NaN / Inf);

[0169] Parameters include the probability of triggering the entire mask. ( ), Maximum number of whole-column masks per sample ( Time slice mask trigger probability ( Minimum time slice mask length Maximum time slice mask length ( ≤ ).

[0170] True missing mask During generation, the naturally missing data locations are marked, and invalid values ​​are filtered out. See the following formula for reference:

[0171]

[0172] During the initialization of the composite mask, the entire column mask and the time slice mask are initialized as all-1 tensors (no mask state): .

[0173] When generating a composite mask, the two composite strategies differ in the sample dimension. Independent sampling, with probability Trigger full column masking: For each sample b, randomly sample the number of columns to be masked, and randomly sample the set of columns. ,satisfy ;Will Set all time steps of sample b and the corresponding positions in column S to 0.

[0174] During time-slice synthesis mask generation, using probability Trigger time slice mask: For each sample b, randomly sample the performance metrics to be masked. ;Will (samples b, s to Set the time step and the corresponding position in column c to 0.

[0175] Final observation mask During composition, real-world missing data and synthetic missing data are combined to align with the real-world scene. .

[0176] Optionally, if the mask tensor of a sample is all 0 ( If ), then one observable bit is randomly reserved as 1 to avoid training failure due to the model having no effective input.

[0177] The time characteristic is always 1: =1 (all 1 tensors, ensuring temporal features are always visible), complete mask concatenation (Concatenate the KPI mask with the time feature mask, with the same dimensions as the input features);

[0178] An encoder architecture that combines data reshaping and parameter sharing can be adopted to transform multi-data feature modeling into parallel processing of multiple single-feature sequences, avoiding numerical interference between heterogeneous indicators and achieving cross-indicator knowledge sharing.

[0179] Input can be made in the rearrangement and shared encoder unit. , , ,in, (Performance data mask section) (The time feature mask is always 1);

[0180] First, mask the performance data, then concatenate it with time features. .

[0181] Each performance data point is concatenated with the same time feature and rearranged into a channel-independent form, where the feature tensor is rearranged as follows: (New batch size = original batch size × number of indicator features, each new sample corresponds to one indicator feature channel of one original sample); the mask tensor is synchronously rearranged as follows: (Depend on and (Synchronous rearrangement, maintaining the correspondence between mask and feature); attention mask generation is... (Attention visibility mask, (Mark the visibility of each new sample time step).

[0182] The rearranged data is fed into a shared temporal encoder (multi-head self-attention + residual feedforward network, encoder dimension d=256), and encoding is performed using... As an attention mask in the time dimension, the final output channel hidden sequence This provides temporal characteristics for subsequent reconstruction and prediction tasks.

[0183] S309. By observing the mask matrix, the standardized performance datasets corresponding to each time period are masked to obtain the masked performance datasets corresponding to each time period.

[0184] Only the standardized performance datasets corresponding to each time period. Perform masking to standardize the target dataset. As a training supervision signal, no masking operation is performed.

[0185] Match the observation mask to each pair of historical-future time periods. (Independent allocation per sample ensures diversity in mask distribution); masking is achieved through element-wise multiplication. After masking, 0 values ​​in the masked performance dataset correspond to "data missing / invisible", and non-zero values ​​correspond to "data valid / visible", which is consistent with the data missing pattern in real operation and maintenance scenarios.

[0186] After masking, it is necessary to verify the effective data ratio (non-zero value ratio) of each masked performance dataset to ensure that the effective data ratio is not less than 10%, so as to avoid the model being unable to learn effective temporal patterns due to excessive masking.

[0187] S310. Based on the mask performance dataset, input time feature vector, standardized target dataset and target time feature vector corresponding to each time period, determine the time period dataset corresponding to each time period.

[0188] For any given time period, construct time period input samples based on the masked performance dataset and the input time feature vector; construct time period prediction samples based on the standardized target dataset and the target time feature vector; and determine the time period input samples and time period prediction samples as the time period dataset.

[0189] The complete input features can be obtained by concatenating dimensions based on the masked performance dataset and the input time feature vector. Simultaneously add the corresponding complete mask. and attention mask This constitutes a complete time-segment input sample (including features, masks, and attention constraints).

[0190] Complete predictive supervision features can be obtained by concatenating dimensions based on a standardized target dataset and a target time feature vector. This sample serves as a supervisory signal for the model's prediction task, used to calculate the prediction loss and guide the model to learn the temporal mapping pattern from history to the future.

[0191] Input samples for different time periods (input feature tensors, ...) , The hidden sequence E after independent channel rearrangement, along with the “time period prediction sample (prediction supervision tensor)”, is used to determine the time period dataset corresponding to each time period.

[0192] The final time period dataset format is as follows: This provides complete input and supervision information for subsequent model reconstruction and prediction.

[0193] After the model is built, it is necessary to verify the dimensionality consistency of each time period dataset (such as matching the input features with the mask dimensions and the continuity of historical and future sequences) and the uniformity of data types (all are floating-point tensors) to avoid model training errors due to data incompatibility.

[0194] S311. By reconstructing the prediction model, the datasets for each time period are corrected to obtain multiple target time period datasets corresponding to multiple time periods.

[0195] S312. Train the initial multi-task model using multiple target time period datasets to obtain the target multi-task model.

[0196] The execution process of S311-S12 can be found in the execution process of S205-S206, and will not be repeated here.

[0197] The multi-task model training method provided in this application can determine the time feature vector corresponding to each time period based on the standardized input value and the standardized predicted value. At the same time, it adopts the channel-independent modeling method to eliminate feature crosstalk between heterogeneous indicators, which can reduce feature crosstalk between performance indicators in different time periods, improve the model's adaptability to data with high missing rate, and thus improve the accuracy of multi-task model training.

[0198] The following are specific embodiments improved from the embodiments of this application.

[0199] Example 1:

[0200] The scenario scale is 30,000 to 50,000 cells in a single province, and the number of performance indicators is K≈60; the training hardware configuration is 2×A100GPU.

[0201] Task 1 (Missing Data Reconstruction): Complete the data at the masked positions within the historical time window of the day (L=96, 15 minutes / step), output KPI (performance index) data consistent with the original physical dimensions, and complete the recovery of historical missing data.

[0202] Task 2 (Next Day Prediction): Based on historical complete / complete data, output the time series prediction results of each KPI for the next H=96 time steps (the whole day of the next day, 15 minutes / step). The output data is consistent with the original physical dimensions.

[0203] Data preprocessing and standardization alignment: First, quality control is performed on the raw KPI data (removing indicator columns with missing rates > 30% and filtering samples with discontinuous timestamps). Then, robust statistics across the entire network (global median and interquartile range IQR) are used in combination with a unified scaling factor s to standardize the data and unify the data scale. Fixed parameters: s = 1.0, ε = 1e−6. The standardization formula and the inverse transformation formula strictly follow the consistent standards of training / inference.

[0204] Time feature construction: 7-dimensional fixed time features are adopted (4-dimensional periodic terms: hourly sine / cosine coding + 15-minute sine / cosine coding; 3-dimensional business time period one-hot coding: working hours, morning and evening peak hours, and nighttime off-peak hours), and the historical sequence and future sequence are seamlessly connected, and the feature construction rules are consistent throughout the process.

[0205] Mask-aware modeling: Generating realistic observation masks based on real missing data in the field. Then, two types of composite masks (the entire column composite mask + the time slice composite mask) are superimposed to obtain the final observation mask. complete mask splicing time feature all-1 mask; fixed parameters: (Probability of triggering the entire mask) (Time slice mask trigger probability) =4 (minimum time slice length) =12 (maximum time slice length), optional robustness processing: random retention of full-mask samples.

[0206] Channel-independent encoding and dual-task decoding: The architecture of "channel-independent encoder + reconstruction head + autoregressive prediction head + column gain correction unit" is adopted to perform channel-by-channel 1D decoding of multiple KPIs; the encoder dimension d=256 (multi-head self-attention + residual feedforward network), and a column-by-column learnable gain parameter g (initialized to 8~16 with range constraints) is applied to the end of the prediction branch to solve the problem of compression of the predicted amplitude of different dimensional indicators.

[0207] Joint loss optimization consists of four components, with default weights: α=0.05 (reconstruction loss Lrec, calculated only at the mask position), γ=0.20 (prediction loss). (covering all 96 future time steps), β=0.0 (temporal consistency loss) Optional constraints include ensuring the last position of the historical sequence matches the first position of the predicted sequence in terms of temporal trend, and λ=0.0 (column gain regularization loss). The optimizer uses AdamW with a gradient pruning threshold of 1.0 and optional EMA (exponential moving average, coefficient = 0.995).

[0208] Training duration: Offline full-data training ≤ 12 hours. After training, save the complete model checkpoint (including model weights, scaler.pkl, temporal feature rules, feature column order features[]). ).

[0209] Example 2: Edge-cloud collaboration (near real-time reconstruction + short-term prediction).

[0210] Deployment configuration: Edge-end performs preprocessing and forward inference (reconstruction + short-term prediction) of I1–I5, and the center performs periodic training and distributes checkpoints.

[0211] Consistency Guarantee: Use and Verify the column order and feature set; scaler.pkl carries {median, IQR, s, cols, version} to ensure consistency with the training caliber; the inverse transformation order and pruning rules are fixed ( / s→inverse→physical pruning).

[0212] Reconstruction task: Fill in the missing positions at the edge in real time;

[0213] Prediction task: The edge generates a rolling prediction of the future H=96 points for alarm suppression and resource scheduling;

[0214] The masking strategy follows the training configuration ( / (span range is consistent), during inference Data from the field isfinite.

[0215] Link Description: Only the inference engine and configuration are retained at the edge, while training and metric monitoring are unified at the central side, which can meet the requirements of cross-region reuse and one-click rollback.

[0216] Example 3: Online Incremental / Small-Step Fine-Tuning

[0217] Applicable scenarios: Slow data distribution drift or the launch of new services causing local metric deviations.

[0218] Daily fine-tuning: Freeze the shared encoder (140), adjust only the learnable gain (153) + reconstruction head (151) + prediction head (152); keep the normalizer and temporal feature construction unchanged;

[0219] Triggering conditions: If the online SMAPE degradation exceeds 5% (or a specific KPI trend drift) is monitored using a 7-day rolling window, a full retraining (unfreezing the encoder) will be triggered.

[0220] Risk control and rollback: All models are versioned in checkpoint form; fine-tuning only takes effect in the "grayscale area," through... and Verification ensures consistency between online and offline versions; in case of anomalies, the system will revert to the previous stable version.

[0221] Benefits: Without changing the core coding capabilities, it can quickly absorb local distribution changes, stabilize the prediction amplitude and alarm accuracy, and reduce the frequency of full retraining and computing power consumption.

[0222] Figure 4 A schematic diagram of the architecture of a multi-task model training method provided in this application. Please refer to [link / reference]. Figure 4 It includes five major modules: data plane, model core, training path, inference path, and control / validation plane, realizing a closed loop of the entire process from data preprocessing and model training to inference service.

[0223] The standardization and feature construction of input data are completed in the data plane. The data plane is the pre-processing stage for model training and inference. When the data is accessed, the original KPI data (including cell identifier, timestamp, and multi-dimensional performance indicators) is received to provide the original data source for subsequent processing. Standardization and alignment (110) is based on the robust statistics of the whole network (median, interquartile range) and the unified scaling factor to convert KPI data from different sources and magnitudes into a unified standardized format to eliminate regional and equipment differences. Mask generation (130) combines real data missing and synthetic mask (entire column / time slice) to generate observation mask to simulate real missing scenarios. Time feature engineering (120) generates 7-dimensional time features (hour / 15-minute period items + business time period one-hot) based on data timestamps to provide time series benchmark for the model.

[0224] The control / validation surface ensures consistency between training and inference. It runs through the entire training and inference process, ensuring unified configuration of data and models. Feature and version configurations can manage core configurations such as performance metric column names and time feature rules. The normalizer parameters can store parameters such as median, interquartile range, and scaling factor required for standardization (i.e., scaler.pkl). Column order locking and validation (161-162) locks the feature column order in the training and inference phases respectively and uses version hash validation to avoid drift in the scope.

[0225] The core of the model includes channel-independent encoding and dual-task decoding capabilities. Channel-independent encoder (140): Concatenates and rearranges standardized KPIs and temporal features into a channel-independent form, and extracts temporal features through a shared multi-head self-attention + residual feedforward network; Autoregressive decoder (152): Learns the temporal mapping from history to the future based on the encoded features and outputs preliminary prediction results; Learnable gain (153): Applies column-wise learnable gain to the prediction results to solve the problem of amplitude compression of weak dimensional indicators; Reconstruction decoder (151): Completes missing data based on encoded features and outputs reconstruction results for historical periods; Joint loss (150): Integrates reconstruction loss, prediction loss and other multi-task losses to provide supervision signals for model parameter updates.

[0226] The training path constructs training samples based on the data surface output, completes model training, and saves the results: Training input assembly: The standardized data, temporal features, and observation masks from the data surface output are assembled into training inputs; Model checkpoints: After training, the model weights, normalizer parameters, feature configurations, and other information are saved to form a self-contained model package.

[0227] The inference path loads training results, processes field data, and outputs business-usable results: Load checkpoints (160): Load the model checkpoints generated during training to restore the model and configuration; Inference input assembly: Perform standardization, time feature construction, and column order verification on the field data and assemble it as inference input; Prediction / reconstruction output assembly: Obtain the prediction / reconstruction results output by the model; Inverse transformation and physical clipping (163): Restore the output of the standardized space to the original physical dimensions and clip it to a reasonable range for business; Service interface API: Provide KPI data services for missing reconstruction and future prediction to the outside world.

[0228] Figure 5 This is a schematic diagram of the architecture of a multi-task model provided in this application. Please refer to [link / reference]. Figure 5 This includes single-channel preprocessing, modeling, and loss supervision. Single-channel preprocessing enables independent input of multiple KPIs. The standardized KPI (single-channel) input is the standardized data of a specific KPI (obtained from the original KPI through robust standardization). The observation mask (single-channel) corresponds to the observation mask for that KPI (1 indicates visible data, 0 indicates missing / reconstruction required). The time feature (historical) reflects the 7-dimensional historical time feature corresponding to that KPI (hourly / 15-minute period items + one-hot data for business periods). The single-channel KPI data is concatenated with the shared time feature to form a basic sequence of "1 KPI + 7-dimensional time feature". The concatenated sequence is then rearranged into a single-channel input format, ultimately outputting an independent input sequence corresponding to a single KPI (this process needs to be repeated for multiple KPIs to form a batch single-channel sequence of (B⋅K)×L×(1+T).

[0229] The model achieves cross-KPI knowledge reuse based on a shared encoder, and supports dual tasks through dual-branch decoding. The shared temporal encoder (140) receives all single-channel input sequences, extracts temporal features through a multi-head self-attention + residual feedforward network, and outputs the hidden sequences of each channel (realizing the sharing of temporal patterns across KPIs); the reconstruction decoder (151) decodes the missing positions of the KPI (regions where the observation mask is 0) time-by-time based on the hidden sequences, and outputs the reconstruction result of the KPI; the autoregressive decoder (152) learns the mapping from historical temporal features to the future based on the same hidden sequence, and outputs the preliminary prediction result of the KPI; the learnable gain (153) applies a channel-by-channel learnable gain to the preliminary prediction result to correct the predicted amplitude of weak-dimensional KPIs (solving the amplitude compression problem caused by the difference in dimensions of different KPIs).

[0230] Loss supervision optimizes the accuracy of both reconstruction and prediction tasks through joint supervision of multiple losses. The reconstruction loss (only at mask positions) calculates the error between the reconstruction result and the true value only for missing positions where the observation mask is 0, supervising the accuracy of missing data completion; the prediction loss (all future steps) calculates the error between the prediction result and the true target value for all future time steps, supervising the prediction accuracy of future time series; the consistency loss is optional and is used to constrain the trend consistency between the reconstruction result of the last position in the historical sequence and the first position in the predicted sequence, improving the continuity of the time series.

[0231] Figure 6 This is a schematic diagram of a multi-task model training device provided in this application. Please refer to [link / reference]. Figure 6 The device 600 may include an acquisition module 601, a first determination module 602, a second determination module 603, a third determination module 604, a correction processing module 605, and a training module 605.

[0232] The acquisition module 601 is used to acquire multiple standardized input values ​​and multiple standardized prediction values ​​corresponding to the initial multi-task model. The standardized input values ​​include multiple standardized performance data corresponding to each performance indicator among multiple performance indicators, and the standardized prediction values ​​include multiple standardized target data corresponding to each performance indicator.

[0233] The first determining module 602 is used to determine the standardized performance dataset and standardized target dataset corresponding to each group of time periods based on the historical time periods of each standardized performance data corresponding to each performance index and the future time periods corresponding to each standardized target data. The group of time periods is the time period in which the historical time period and the future time period have a temporal correlation relationship.

[0234] The second determining module 603 is used to determine the input time feature vector of the standardized performance dataset corresponding to each time period and the target time feature vector of the standardized target dataset based on the preset multi-dimensional time features.

[0235] The third determining module 604 is used to determine multiple time period datasets corresponding to multiple groups of time periods based on the standardized performance dataset, input time feature vector, standardized target dataset and target time feature vector corresponding to each group of time periods.

[0236] The correction processing module 605 is used to correct the datasets of each time period by reconstructing the prediction model, so as to obtain multiple target time period datasets corresponding to multiple groups of time periods.

[0237] The training module 605 is used to train the initial multi-task model using multiple target time period datasets to obtain the target multi-task model.

[0238] In one possible implementation, the second determining module 603 is specifically used for:

[0239] Obtain multiple time feature rules corresponding to preset multidimensional time features;

[0240] By using multiple time feature rules, the time series information of the standardized performance dataset corresponding to each historical period is processed to extract features, and the input time feature vector corresponding to each historical period is obtained.

[0241] By using multiple time feature rules, the time series information of the standardized target dataset corresponding to each future time period is processed to extract features, and the target time feature vector corresponding to each future time period is obtained.

[0242] In one possible implementation, the multiple time feature rules include multiple periodic item rules and multiple business time period rules; for any given time period; the second determining module 603 is specifically used for:

[0243] Based on multiple periodic rules, feature processing is performed on time periods to obtain time feature values ​​corresponding to each periodic rule.

[0244] Based on multiple business time period rules, feature processing is performed on the time periods to obtain the business time period feature values ​​corresponding to each business time period rule;

[0245] The time feature values ​​corresponding to each periodic item rule and the business time period feature values ​​corresponding to each business time period rule are concatenated to obtain the input time feature vector corresponding to the standardized performance data.

[0246] In one possible implementation, the third determining module 604 is specifically used for:

[0247] Generate an observation mask matrix, which is used to indicate the mask position in the normalized performance dataset corresponding to each time period;

[0248] By observing the mask matrix, the standardized performance datasets corresponding to each time period are masked to obtain the masked performance datasets corresponding to each time period.

[0249] Based on the mask performance dataset, input time feature vector, standardized target dataset, and target time feature vector corresponding to each time period, the time period dataset corresponding to each time period is determined.

[0250] In one possible implementation, for any given time period; the third determining module 604 is specifically used for:

[0251] Based on the mask performance dataset and the input time feature vector, construct the time period input sample;

[0252] Based on the standardized target dataset and target time feature vector, construct time period prediction samples;

[0253] The input samples and predicted samples for each time period are used to define the time period dataset.

[0254] In one possible implementation, the reconstruction prediction model includes a channel-independent encoder, a reconstruction head, an autoregressive prediction head, a column gain correction unit, and a joint loss calculation unit.

[0255] In one possible implementation, the acquisition module 601 is specifically used for:

[0256] Obtain multiple initial input values ​​corresponding to multiple standardized input values ​​and multiple initial prediction values ​​corresponding to multiple standardized prediction values;

[0257] Determine multiple intermediate input values ​​corresponding to multiple initial input values ​​and multiple intermediate prediction values ​​corresponding to multiple initial prediction values;

[0258] Multiple intermediate input values ​​and multiple intermediate predicted values ​​are scaled separately to obtain multiple standardized input values ​​and multiple standardized predicted values.

[0259] The multi-task model training device provided in this application embodiment can execute the technical solution shown in the above method embodiment. Its implementation principle and beneficial effects are similar, and will not be described again here.

[0260] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Please refer to... Figure 7 The electronic device 700 may include at least one processor 701 and a memory 702. Exemplarily, the processor 701 and the memory 702 are interconnected via a bus 703.

[0261] Memory 702 stores instructions executed by the computer;

[0262] The processor 701 executes computer execution instructions stored in the memory 702, causing the processor 701 to perform the test method as shown in the above method embodiment.

[0263] The specific implementation process of processor 701 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0264] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0265] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0266] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0267] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0268] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0269] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0270] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0271] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, devices, or units, and may be electrical, mechanical, or other forms.

[0272] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0273] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0274] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0275] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0276] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A multi-task model training method, characterized in that, include: Obtain multiple standardized input values ​​and multiple standardized predicted values ​​corresponding to the initial multi-task model. The standardized input values ​​include multiple standardized performance data corresponding to each performance indicator among multiple performance indicators, and the standardized predicted values ​​include multiple standardized target data corresponding to each performance indicator. Based on the historical time periods corresponding to each standardized performance data and the future time periods corresponding to each standardized target data, the standardized performance dataset and standardized target dataset corresponding to each group of time periods are determined. The group of time periods is the time period in which the historical time period and the future time period have a temporal correlation relationship. Based on preset multidimensional time features, the input time feature vector of the standardized performance dataset corresponding to each group of time periods and the target time feature vector of the standardized target dataset are determined. Based on the standardized performance dataset, input time feature vector, standardized target dataset and target time feature vector corresponding to each group of time periods, determine multiple time period datasets corresponding to the multiple groups of time periods; By reconstructing the prediction model and correcting the datasets for each time period, multiple target time period datasets corresponding to the multiple time periods are obtained. The initial multi-task model is then trained using the multiple target time period datasets to obtain the target multi-task model.

2. The method according to claim 1, characterized in that, Based on preset multidimensional time features, the input time feature vector of the standardized performance dataset corresponding to each group of time periods and the target time feature vector of the standardized target dataset are determined, including: Obtain multiple time feature rules corresponding to the preset multidimensional time features; By using the aforementioned multiple time feature rules, the time series information of the standardized performance dataset corresponding to each historical period is processed for feature extraction to obtain the input time feature vector corresponding to each historical period. By using the aforementioned multiple time feature rules, the time series information of the standardized target dataset corresponding to each future time period is processed to extract features, thereby obtaining the target time feature vector corresponding to each future time period.

3. The method according to claim 2, characterized in that, The multiple time feature rules include multiple periodic item rules and multiple business time period rules; for any given time period; through the multiple time feature rules, feature extraction processing is performed on the standardized performance dataset corresponding to the time period to obtain the input time feature vector corresponding to the time period, including: Based on the multiple periodic item rules, feature processing is performed on the time period to obtain the time feature value corresponding to each periodic item rule; Based on the multiple business time period rules, feature processing is performed on the time periods to obtain the business time period feature values ​​corresponding to each business time period rule; The time feature values ​​corresponding to each periodic item rule and the business time period feature values ​​corresponding to each business time period rule are concatenated to obtain the input time feature vector corresponding to the standardized performance data.

4. The method according to claim 1, characterized in that, Based on the standardized performance dataset, input time feature vector, standardized target dataset, and target time feature vector corresponding to each group of time periods, multiple time period datasets corresponding to the multiple groups of time periods are determined, including: Generate an observation mask matrix, which is used to indicate the mask position in the standardized performance dataset corresponding to each time period; The standardized performance datasets corresponding to each time period are masked using the observation mask matrix to obtain the masked performance datasets corresponding to each time period. Based on the mask performance dataset, input time feature vector, standardized target dataset, and target time feature vector corresponding to each group of time periods, the time period dataset corresponding to each group of time periods is determined.

5. The method according to claim 4, characterized in that, For any given time period group; based on the masked performance dataset, input time feature vector, standardized target dataset, and target time feature vector corresponding to the time period group, determine the time period dataset corresponding to the time period, including: Based on the mask performance dataset and the input time feature vector, construct time period input samples; Based on the standardized target dataset and the target time feature vector, construct time period prediction samples; The input samples and predicted samples for the specified time period are used to determine the time period dataset.

6. The method according to claim 4, characterized in that, The reconstruction prediction model includes a channel-independent encoder, a reconstruction head, an autoregressive prediction head, a column gain correction unit, and a joint loss calculation unit.

7. The method according to claim 1, characterized in that, Obtain multiple standardized input values ​​and multiple standardized prediction values ​​corresponding to the initial multi-task model, including: Obtain multiple initial input values ​​corresponding to multiple standardized input values ​​and multiple initial prediction values ​​corresponding to the multiple standardized prediction values; Determine multiple intermediate input values ​​corresponding to the multiple initial input values ​​and multiple intermediate prediction values ​​corresponding to the multiple initial prediction values; The intermediate input values ​​and the intermediate predicted values ​​are scaled to obtain the standardized input values ​​and the standardized predicted values.

8. A multi-task model training device, characterized in that, It includes an acquisition module, a first determination module, a second determination module, a third determination module, a correction processing module, and a training module: The acquisition module is used to acquire multiple standardized input values ​​and multiple standardized prediction values ​​corresponding to the initial multi-task model. The standardized input values ​​include multiple standardized performance data corresponding to each performance indicator among multiple performance indicators, and the standardized prediction values ​​include multiple standardized target data corresponding to each performance indicator. The first determining module is used to determine the standardized performance dataset and standardized target dataset corresponding to each group of time periods based on the historical time periods corresponding to each standardized performance data and the future time periods corresponding to each standardized target data, wherein the group of time periods is a time period in which the historical time period and the future time period have a temporal correlation relationship; The second determining module is used to determine, based on preset multi-dimensional time features, the input time feature vector of the standardized performance dataset corresponding to each group of time periods, and the target time feature vector of the standardized target dataset; The third determining module is used to determine multiple time period datasets corresponding to the multiple groups of time periods based on the standardized performance dataset, input time feature vector, standardized target dataset and target time feature vector corresponding to each group of time periods; The correction processing module is used to correct the datasets of each time period by reconstructing the prediction model, so as to obtain multiple target time period datasets corresponding to the multiple groups of time periods. The training module is used to train the initial multi-task model using the multiple target time period datasets to obtain the target multi-task model.

9. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-7.

Images